docs(mcts): explain unconditional model call on done lanes#18
Merged
Conversation
The recurrent_fn passed to mctx evaluates the model on every batch lane, even where the env has terminated. This was flagged in #15 as potential wasted compute and a NaN risk, but on inspection the behavior is structurally correct: - mctx's search loop calls recurrent_fn with a fixed-shape batch. - Under jit + vmap, the model forward is a single XLA op that runs on the whole batch; per-lane skipping would need dynamic shapes. - Done-lane output is algorithmically inert: value is masked to 0 on the next line and discount is also 0, so terminal lanes contribute nothing to mctx's backup. Replace the open "deferred to a focused PR" promise with a comment that captures the rationale so future readers don't re-litigate this dead end. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Closes the open promise from #15 ("MCTS terminal-state model eval ... deferred to a focused PR") — with an honest "we looked and it doesn't need fixing" rather than a code change.
The recurrent_fn passed to mctx evaluates the model on every batch lane, including ones where the env has terminated. On closer inspection:
recurrent_fnwith a fixed-shape batch. Under jit + vmap, the model forward is a single XLA op that runs on the whole batch; per-lane skipping would require dynamic shapes, which jit forbids.valueis masked to 0 on the next line anddiscountis also 0, so terminal lanes contribute nothing to mctx's backup.So the unconditional model call is structurally correct, not a bug. Adding defensive zeros-substitution would violate "don't add validation for scenarios that can't happen" and add code with no algorithmic effect.
This PR adds a comment block at the top of
recurrent_fncapturing the rationale so future readers (and future-me) don't re-litigate the question.Test plan
🤖 Generated with Claude Code